Frontier Artificial Intelligence Models

A frontier AI model is a highly capable, general-purpose model whose performance is at or near the leading edge of existing artificial intelligence. It is defined by a combination of broad competence and the potential to develop dangerous capabilities, rather than by one fixed parameter count or benchmark score. Because the technological frontier advances, the classification is relative and changes over time.

Core technological capabilities

Frontier status depends on demonstrated or reasonably foreseeable performance across several domains.

• The model can perform a wide range of tasks, including language generation, coding, reasoning, knowledge retrieval and, for multimodal models, processing more than one type of data.
• It can adapt to unfamiliar tasks through prompting, in-context learning or fine-tuning, reducing the need for task-specific models.
• It may support multi-step planning, tool use and increasingly autonomous action when connected to software, external data or other systems.
• It can achieve broad performance near the state of the art, rather than merely excelling at one narrow task.

Capabilities relevant to severe harm

A model becomes especially significant when evaluation indicates that it could materially assist activities capable of causing large-scale harm.

• Relevant capabilities include advanced assistance in cyber operations, biological or chemical misuse, and the evasion of safeguards.
• Other concerns include deception, manipulation, autonomous replication or adaptation, and assistance with developing more capable AI systems.
• Risk depends on both the model and its deployment environment: tool access, memory, system prompts and agentic scaffolding can substantially increase effective capability.

How frontier capability is assessed

Developers and regulators use capability evaluations, red-teaming and domain-expert testing to examine both ordinary performance and hazardous capabilities. Training compute, model size and data scale are useful proxies, but they do not alone establish frontier status because architecture, training methods and system integration also affect performance.